Necessary file of running customized YOLOv3

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hi,

I’m working with deepstream-test5 with kafka. The default AI model of ds_test5 is resnet10. Due to lack of accuracy, I changed it to YoloV3, and worked perfectly.

The files below are those which been used in config file. In my project, I may need to train my own Yolo network but I’m not sure which file need to be replaced?

1. yolov3.cfg
2. yolov3.weights
3. labels.txt
4. yolov3-calibration.table.trt5.1
5. libnvdsinfer_custom_impl_Yolo.so

I can generate 1~3 by myself, but I’m not sure what is 4 and 5. Do I need to replace them if I want to use my own yolov3 network? If so, how can I generate them?

3

Hi,

  1. Is the calibration table required to generate an INT8 model. Currently you will either need to generate it yourself in a stand alone tool or you can modify the nvinfer plugin source code. Please refer to this answer for more details.

  2. Is the custom library which generates the yolo n/w arch in TRT and also parse the output bounding boxes. You can refer to the source code in samples\Objectdetector_Yolo\ directory. Please note the src code in the SDK only supports 4 standard yolo models and the tlt-yolo model. You will need modify these sources to adapt to your custom model. Please refer to this FAQ doc from DS 4.0.2 for some details on what changes are usually required.